AI systems that explain their decisions can be monitored for harmful intentions.
problem Monitoring AI systems' decision-making processes for harmful intentions is imperfect and can miss some misbehavior.
method Monitoring the chain of thought (CoT) of AI systems that communicate in human language.
result CoT monitoring is a promising but fragile approach to AI safety.
Simple online monitor detects unsafe LLM outputs.
problem LLMs generate unsafe outputs despite training.
method Thresholding external verifier signal to decide alarms.
result Simple design competitive with advanced methods.
SafeML monitors ML systems for safety and security risks.
problem Ensuring safety and explainability of ML systems in safety-critical domains.
method Statistical difference measures of ECDF to detect distributional shifts.
result Approach can detect invalid application contexts of ML components.
Real time large scale streaming data pose major challenges to forecasting, in particular defying the presence of human experts to perform the corresponding analysis. We present here a class of models and methods used to develop an automated, scalable and versatile system for large scale forecasting oriented towards saf…
System predicts ice formation to improve road safety.
problem Improving road safety through accurate ice formation prediction.
method Combining physical models and machine learning for data correction.
result Machine learning effectively corrects incorrect data in forecasts.
Paper monitors DNN accuracy to enhance trustworthiness.
problem Varying DNN accuracy in practice and lack of ground truth labels.
method Post-hoc accuracy monitor model using Monte-Carlo dropout ensemble.
result Accuracy monitor provides close-to-true accuracy estimation.
Online monitoring system for safety classifiers with shift detection and conformal adaptation
problem Detecting and adapting to distributional shifts in deployed safety classifiers
method Calibrated sequential statistics for online monitoring, conformal abstention for adaptation
result 86.6% valid detection with mean latency of 39.5 steps
Online monitor detects classifier drift and adapts predictions.
problem Silent degradation of classifier accuracy under distributional shift.
method Sliding-window KS statistic with calibrated alarm thresholds.
result 86.6% valid detection across various shift conditions.
For using neural networks in safety critical domains, it is important to know if a decision made by a neural network is supported by prior similarities in training. We propose runtime neuron activation pattern monitoring - after the standard training process, one creates a monitor by feeding the training data to the ne…
Proposes real-time risk monitoring for machine learning systems under unknown shifts.
problem Dynamic distribution shifts challenge real-world machine learning systems' risk assurances.
method Sequential hypothesis testing with 'testing by betting' to detect risk violations.
result Effective real-time risk monitoring under various unknown shifts.
AI monitors social distancing and masks at manufacturing plants.
problem Ensuring safety of workers during post-COVID production.
method Computer vision and AI techniques for social distancing and mask detection.
result Real-time alerts prevent violations of social distancing and mask-wearing.
The paper proposes an on-line monitoring framework for continuous real-time safety/security in learning-based control systems (specifically application to a unmanned ground vehicle). We monitor validity of mappings from sensor inputs to actuator commands, controller-focused anomaly detection (CFAM), and from actuator c…
The paper proposes a method to create efficient remote monitoring models.
problem Large and complex machine learning models are unsuitable for remote monitoring on edge devices.
method Decompose the model into a simple local monitoring function and a complex correction term evaluated on the server.
result The proposed framework learns monitoring models with significantly reduced complexity that maintain safety.
This research tackles monitoring machine learning algorithms post-deployment, addressing performativity issues.
problem Monitoring machine learning algorithms after deployment, especially when they affect their own data-generating process.
method Uses causal inference techniques to navigate performativity and compares different monitoring criteria and data sources.
result Different monitoring systems have varying operating characteristics and implications for ML monitoring design.
The failure of a complex and safety critical industrial asset can have extremely high consequences. Close monitoring for early detection of abnormal system conditions is therefore required. Data-driven solutions to this problem have been limited for two reasons: First, safety critical assets are designed and maintained…
Although aviation accidents are rare, safety incidents occur more frequently and require a careful analysis to detect and mitigate risks in a timely manner. Analyzing safety incidents using operational data and producing event-based explanations is invaluable to airline companies as well as to governing organizations s…
New approach extracts AI model representations for steering and monitoring.
problem Understanding and steering AI models' internal knowledge representation.
method Feature learning advances for scalable linear concept representations.
result Concept representations enable steering and monitoring of AI models.
Novel method detects faults in helicopter transmissions using healthy data only.
problem Detecting faults in rare events in helicopter transmissions.
method Probabilistic anomaly detection using Bayesian measures.
result Method achieves competitive detection performance in real-world and benchmark applications.
In multimodal traffic monitoring, we gather traffic statistics for distinct transportation modes, such as pedestrians, cars and bicycles, in order to analyze and improve people's daily mobility in terms of safety and convenience. On account of its robustness to bad light and adverse weather conditions, and inherent spe…
Can engineering neural networks be approached in a disciplined way similar to how engineers build software for civil aircraft? We present nn-dependability-kit, an open-source toolbox to support safety engineering of neural networks for autonomous driving systems. The rationale behind nn-dependability-kit is to consider…
This paper reviews traditional and modern methods for detecting structural damage using vibrations.
problem Early warning of structural damage to maintain civil structures.
method Vibration-based methods and ML/DL algorithms.
result ML and DL algorithms show superior performance in detecting structural damage.
Efficient classifier with uncertainty bounds for safety-critical applications.
problem Lack of uncertainty bounds in high-accuracy classifiers for safety-critical tasks.
method Nadaraya-Watson estimator with frequentist bounds.
result Competitive accuracy and uncertainty bounds at reduced computational cost.
Fine particulate matter (PM2.5) is one of the criteria air pollutants regulated by the Environmental Protection Agency in the United States. There is strong evidence that ambient exposure to (PM2.5) increases risk of mortality and hospitalization. Large scale epidemiological studies on the health effects of P…
In a world of global trading, maritime safety, security and efficiency are crucial issues. We propose a multi-task deep learning framework for vessel monitoring using Automatic Identification System (AIS) data streams. We combine recurrent neural networks with latent variable modeling and an embedding of AIS messages t…
New framework uses OR to ensure AI systems make safe decisions.
problem Ensuring generative AI systems make safe decisions as they gain autonomy.
method Developed a conceptual framework combining flow-based models and adversarial robustness.
result Increased autonomy requires new OR approaches for feasibility, robustness, and stress testing.
Deep learning identifies precipitation clouds from all-sky camera data.
problem Automating cloud warning systems for observatories.
method Deep learning using EfficientNet network.
result Average accuracy of 99% in identifying rainfall potential and 96% in cloud coverage.
Despite the tremendous advances that have been made in the last decade on developing useful machine-learning applications, their wider adoption has been hindered by the lack of strong assurance guarantees that can be made about their behavior. In this paper, we consider how formal verification techniques developed for …
Transformer model predicts train axle vibrations for safer maintenance.
problem Prevent mechanical failures in railway axles.
method Integrates Deep Autoregressive solution with spectral methods and observation models.
result Transformer model (ShaftFormer) improves predictive maintenance for railway axles.
Study optimizes inspection and monitoring of deteriorating structures using POMDPs.
problem Optimizing decision-making for deteriorating engineering systems with uncertain observations.
method Developed theoretical and computational foundations for Value of Information (VoI) and Value of Structural Health Monitoring (VoSHM) in POMDPs.
result POMDP policies inherently leverage VoI to guide optimal observational actions, improving long-term cost.
Using raw sensor data to model and train networks for Human Activity Recognition can be used in many different applications, from fitness tracking to safety monitoring applications. These models can be easily extended to be trained with different data sources for increased accuracies or an extension of classifications …
Real-time detection of out-of-distribution data in CPS control systems.
problem Detecting out-of-distribution data in CPS control systems for safety.
method Inductive conformal prediction and anomaly detection using variational autoencoders and deep support vector data description.
result Efficient real-time detection with low false alarm rates and comparable execution time.
Measures policy-violating content prevalence with ML-assisted sampling and LLM labeling.
problem Accurate measurement of content violations that are often rare and costly to label.
method Design-based measurement system using ML-assisted probability sampling and LLM labeling.
result Produces unbiased prevalence estimates with confidence intervals and dashboard drilldowns.
Gaussian Process Regression improves damage assessment in structural health monitoring.
problem Inaccurate damage quantification in varying conditions and environments.
method Gaussian Process Regression models with Damage Indices for probabilistic damage assessment.
result The method provides probability-based decision-making for structural health monitoring.
FedVision uses federated learning to improve object detection without transmitting data.
problem Challenges in building object detection models on large training datasets due to privacy and cost issues.
method Federated learning (FL) platform for easy integration by non-experts.
result Significant efficiency improvement and cost reduction in smart city applications.
Air traffic control is a real-time safety-critical decision making process in highly dynamic and stochastic environments. In today's aviation practice, a human air traffic controller monitors and directs many aircraft flying through its designated airspace sector. With the fast growing air traffic complexity in traditi…
RLVR maintains safety while improving reasoning capabilities in LLMs.
problem Safety-capability tradeoff in fine-tuning LLMs.
method Reinforcement Learning with Verifiable Rewards (RLVR) and theoretical analysis.
result RLVR can enhance reasoning while maintaining safety guardrails.
Conformal Test Martingales can be 'blind' to significant changes in data distribution.
problem The converse of exchangeability does not hold, leading to potential blindness of CTMs.
method Explicit construction of A-cryptic change-point using bivariate Gaussian distributions. result CTMs can be perfectly cryptic to a significant change in marginal means.
Research tackles safety of deep learning in safety-critical tasks.
problem Safety concerns of deep learning in perception tasks for autonomous agents.
method Technical enumeration and discussions on safety concerns and mitigation methods.
result Need for more mitigation methods to ensure safety of deep learning.
In recent years, deep learning methods have outperformed other methods in image recognition. This has fostered imagination of potential application of deep learning technology including safety relevant applications like the interpretation of medical images or autonomous driving. The passage from assistance of a human d…
Framework predicts remaining useful life of DSH subsystems under unknown failure modes.
problem Predicting remaining useful life of DSH subsystems with unknown failure modes.
method Unsupervised framework using mixture of Gaussian regressions and Expectation-Maximization algorithm.
result Improved prediction accuracy and interpretability of RUL.
How can we find patterns and anomalies in a tensor, or multi-dimensional array, in an efficient and directly interpretable way? How can we do this in an online environment, where a new tensor arrives each time step? Finding patterns and anomalies in a tensor is a crucial problem with many applications, including buildi…
The paper discusses safety assessment for AI systems, focusing on machine learning models.
problem Safety assessment of AI systems, especially machine learning models, in safety-related applications.
method Analyzed AI models as statistical models and proposed a new budget allocation for AI safety.
result Safety assessment of AI systems requires a new approach focusing on the model used, not just the system.
Approves updates to machine learning models in healthcare based on accumulating data.
problem Designing policies to autonomously approve updates to machine learning algorithms in non-stationary settings.
method Learning-to-approve (L2A) approach that uses accumulating monitoring data to learn how to approve modifications.
result L2A learns to abstain when performance drops are common and approves beneficial modifications quickly when the distribution is stable.
Reward hacking exploits misspecified rewards, affecting agent capabilities and true performance.
problem Reward hacking in RL models exploiting reward misspecifications.
method Constructed four RL environments with misspecified rewards; analyzed agent capabilities and behavior.
result More capable agents exploit reward misspecifications, achieving higher proxy reward but lower true reward.
Autonomous vehicles rely on machine learning to solve challenging tasks in perception and motion planning. However, automotive software safety standards have not fully evolved to address the challenges of machine learning safety such as interpretability, verification, and performance limitations. In this paper, we revi…
Fine-tuning LLMs improves capability but harms safety, study finds.
problem Balancing capability and safety in LLM fine-tuning.
method Theoretical framework and numerical experiments for two safety-aware fine-tuning strategies.
result Characterization of fundamental limits of safety-capability trade-off in LLM fine-tuning.
RAGuard improves safety in LLMs for offshore wind maintenance.
problem Conventional LLMs fail with specialised or unexpected scenarios in offshore wind maintenance.
method Integrates safety-critical documents alongside technical manuals in RAG framework.
result RAGuard increases safety recall from almost 0% to over 50% while maintaining technical recall above 60%.
Survey of algorithms for testing AI-driven CPS safety.
problem Testing AI-driven CPS for safety in complex environments.
method Survey of applied algorithms for safety validation.
result Survey of existing tools and techniques for safety validation.